Google Ads MCP: The Official Server, Its Limits, Options

MCP, the Model Context Protocol, is the standard an AI assistant uses to call tools outside itself. An MCP server wraps an API in tools the assistant can discover and call. Google's server wraps the Google Ads API: an assistant connected to it can list the accounts you have access to and run queries against them, then answer in plain language.

The Google Ads API team explored the idea in July 2025 and released the server on GitHub in October 2025, according to trade press coverage. Google's documentation page for it sits in the API's developer toolkit and was last updated on 16 October 2026. It is a developer tool, not a product inside Google Ads: nothing in the Google Ads interface turns it on, and Claude's connector directory lists no ads connector published by Google. You install it, or deploy it, yourself.

• list_accessible_customersReturns the Google Ads accounts the signed-in user or service account can reach. The first call in almost every conversation.
• searchRuns a Google Ads Query Language (GAQL) query: metrics, budgets, statuses, segments. Everything the assistant reports comes through this one tool.
• get_resource_metadataDescribes a resource's fields, so the assistant can write a valid query instead of guessing field names.

The repo also ships resources the assistant can read: the API's discovery document, lists of metrics and segments, and the release notes. In practice that means a good assistant writes correct GAQL on the first or second attempt, because it can look the field up.

Google's documentation is direct about it: the current release is read-only. The assistant can see everything a GAQL query can see and change nothing. For an audit, a weekly report or a question about last month, that is a feature, not a limit. Nobody has to worry that a loosely worded prompt paused the brand campaign.

The keyword ideas limit, as we read it

This one is our reading of the documentation, not a sentence Google wrote. The only data tool is the GAQL search. Keyword ideas and search volumes come from the API's KeywordPlanIdeaService, which is a separate call rather than a resource GAQL can query, and neither the documentation page nor the README mentions it. So, as far as we can tell, the official server cannot return keyword ideas. If keyword research is the job, use the Keyword Planner in Google Ads, or a third-party server that documents it.

There are two ways to run the server, and the choice decides which assistants can reach it. A lot of the "google ads mcp server url" searches end here: there is no URL until you deploy one.

• Locally, over stdio

  Started on your machine with pipx and spoken to through standard input and output. Right for Claude Code, Codex, Cursor or Gemini CLI on the same computer. Nothing is exposed to the internet.

• On your Cloud Run, over HTTP

  Deployed to Google Cloud Run with streamable HTTP. The endpoint is your Cloud Run URL plus /mcp. Right for Claude on the web and ChatGPT, which connect to a remote server over OAuth.

• Google-hosted pilot

  The documentation mentions a hosted endpoint pilot for allowlisted participants. It is not public, and Google does not say on the page how to join.

Clients that do not speak OAuth get an HTTP 401 from the Cloud Run deployment, and Google's advice for them is the local stdio setup. Rate limits for the server itself are not stated; it calls the Google Ads API, so the normal API quotas apply to every query the assistant runs.

The server authenticates the way other Google Cloud tools do, through Application Default Credentials: either OAuth 2.0 as a user, or a service account. The scope is https://www.googleapis.com/auth/adwords. Two environment variables matter.

• GOOGLE_PROJECT_ID is required: the Google Cloud project the server runs under.
• GOOGLE_ADS_LOGIN_CUSTOMER_ID is needed when you reach client accounts through a manager account, as most agencies do. It is the manager's customer ID.

The docs and the README disagree about the developer token

As of October 2026 the two official sources say different things. Google's documentation page says that GOOGLE_ADS_DEVELOPER_TOKEN "was sunset on September 9, 2026", and that the latest version ties the API access level (Explorer, Basic or Standard) to the Google Cloud project instead. The GitHub README still says developer tokens need at least Explorer access. The documentation page is the more recently updated of the two, so start from it, and if a setup fails on a token error, check which version of the server you installed.

Which route you take follows from where the server runs. Google documents Claude and ChatGPT as custom connectors over OAuth on a self-hosted /mcp endpoint, and lists their redirect URIs (claude.ai and chatgpt.com) for that setup. The repo adds Claude Code, Codex, Cursor and VS Code Copilot.

Claude Code, against a Cloud Run deployment

Claude Code adds a remote server with one command. Replace the URL with your own Cloud Run service, then sign in with /mcp inside Claude Code.

claude mcp add --transport http google-ads https://<your-cloud-run-url>/mcp

Claude Code, against a local server

For stdio, Claude Code takes the command that starts the server after a double dash. Use the exact pipx command and environment variables from the repo's README, since they change between versions.

claude mcp add google-ads -- <command that starts the server, from the README>

Claude Code's scopes decide who sees the server: local (the default, only you in this project), project (written to .mcp.json and shared with the repo) or user (you, in every project). For an agency team, project scope keeps one definition in version control.

Claude on the web and Claude Desktop

1. Deploy the server to Cloud Run and note the URL plus /mcp.
2. In Claude, open Settings, then Connectors, and add a custom connector with that URL. On Team and Enterprise an Owner adds it in the organisation settings and members connect individually.
3. If your deployment needs its own OAuth client, enter the client ID and secret under the connector's advanced settings.
4. Connect, sign in with Google, and enable the connector in a chat.

Custom connectors are available on every Claude plan, with Free limited to one. Anthropic notes that it has not verified third-party servers, and to Claude a server you deploy yourself is one of those. ChatGPT works the same way in developer mode: add the server URL as a custom app with OAuth. The setup and the prompts worth asking once it is connected are in Claude for Google Ads.

Next to the server, Google publishes agent skills for the Google Ads API: instructions an assistant loads when a task needs them. They install with one command, and the documentation names three of them: an MCP setup skill, an account diagnostics skill and an API quickstart.

npx skills add google/skills/skills/ads

Google also lists a "Google Ads API Developer Assistant" plugin for Antigravity and Claude Code. The difference matters when you read about a "Claude Google Ads skill": a skill teaches the assistant a procedure, the MCP server gives it the data. You need the server for the skill to have anything to work on.

If you need an assistant that can act on Google Ads, the official server is not it today. Hosted third-party services fill that gap. They are run by companies other than Google, they hold your Google Ads access on their side, and their prices and features move, so treat the table as a starting list as of October 2026, not a verdict.

Three questions to ask any of them before you connect a client account: which OAuth scopes it requests and where the token is stored, whether a write needs your confirmation in the chat or happens on the first instruction, and whether there is a log of what the assistant changed. A read-only server needs none of those answers, which is the official one's quiet advantage. The wider comparison, including Meta, Microsoft and TikTok, is in Ads MCP servers compared, and the trade-offs between an MCP server, a direct API script and Zapier in MCP vs API vs Zapier for ad reporting.

Every Google Ads server, official or not, reads Google Ads. And Google Ads knows what Google Ads measured: clicks, costs, and the conversions its tags and imports recorded inside their conversion windows. For a lead business that conversion is usually a form fill. Whether the lead was qualified, whether it booked a call, whether it closed and for how much: that lives in your CRM, which Google Ads only knows about if you send it back.

So the question an assistant answers best through a Google Ads server is "what does Google Ads say", not "which campaign made us money". Ask Claude for the cheapest campaign per lead and it will find one correctly by Google's numbers, and it may be the one whose leads never close. The worked example is in Why your ad platform's MCP gives Claude the wrong CPL.

• Send the outcome back to GoogleImport qualified leads and won deals as offline conversions or enhanced conversions for leads, so Google's own columns include them. See offline conversion tracking.
• Give the assistant the CRM side directlyConnect a second MCP server that holds the closed revenue per campaign, joined to the click, and ask both in one chat.

LeadJourney tracks every click, form fill and call server-side and first-party, at 95%+ accuracy, and keeps the gclid on the visitor. The lead is followed into the CRM (HubSpot, Salesforce, Pipedrive, Close, Attio, GoHighLevel, ActiveCampaign, Odoo, Zoho CRM and Microsoft Dynamics 365 are native), so each Google Ads campaign, ad group and ad carries its qualified leads, booked demos and won revenue next to the spend the Google Ads integration imports. The CRM stages you map go back to Google Ads as conversions.

Its MCP server is one endpoint, https://api.leadjourney.io/mcp, with OAuth per application, and works with Claude, ChatGPT, Gemini, Copilot and Perplexity. It reads journeys, spend by platform and campaign, and CRM outcomes. What it changes stays inside LeadJourney: reports and dashboards it builds and saves, the attribution model and lookback window. It cannot pause a campaign, move budget or create an ad, and it acts only when you send a message.

That makes the two servers complementary rather than competing. Google's server answers what Google Ads saw; LeadJourney's answers what the CRM says came of it. Where a change follows, you make it in Google Ads, or through a write-capable server you trust. The MCP connection is included from the Growth plan, and below it is an add-on at €49 a month for advertisers or €99 a month flat for agencies. Setup for Claude is on LeadJourney for Claude.

• Ads MCP servers compared: Meta, Google, LinkedIn, TikTok and more
• Claude for Google Ads: connect it, then ask better questions
• Why your ad platform's MCP gives Claude the wrong CPL
• MCP vs API vs Zapier for ad reporting
• Automated budget rules on cost per qualified lead
• Meta Ads MCP: setting up Meta's official server
• Google Ads enhanced conversions for leads
• The LeadJourney MCP server
• The Google Ads integration